update

[lttv.git] / markers-test / test-mark-speed.c

diff --git a/markers-test/test-mark-speed.c b/markers-test/test-mark-speed.c
index b3294e58a781de8cf85e9bf49598856dc72cf400..bd798b8e3de39eda060cec129b48f0ff69aaa9aa 100644 (file)
--- a/markers-test/test-mark-speed.c
+++ b/markers-test/test-mark-speed.c
@@ -10,6 +10,25 @@
 #include <asm/ptrace.h>
 #include <asm/system.h>
 
+static void pmc_flush_cache(void)
+  {
+    /* write back and invalidate cache (a serializing instruction) */
+
+    __asm__ __volatile__ ( "wbinvd" : : : "memory" );
+
+    /* The wbinvd instruction does not wait for the external caches
+     * to be flushed, but only requests that it be done.  The loop
+     * is to be sure that enough time has elapsed, but the compiler
+     * might simplify or even remove it.  The loop bound is for a
+     * 512 KB L2 cache.  On a Pentium Pro/II/III, the loop uses
+     * 2 cycles per iteration.
+     *
+     * Does wbinvd also cause the TLB to be flushed?
+     * A comment in mtrr.c suggests that it does.
+     */
+    { register int i; for (i = 0; i < 512*1024; i++) { } }
+  }
+
 static void noinline test2(const struct marker *mdata,
         void *call_private, ...)
 {
@@ -52,13 +71,23 @@ static void noinline test2(const struct marker *mdata,
        //asm volatile ("");
 struct proc_dir_entry *pentry = NULL;
 
+char temp0[8192];
+int temp[8192] __cacheline_aligned;
+char temp5[8192];
+
 static inline void test(unsigned long arg, unsigned long arg2)
 {
 #ifdef CACHEFLUSH
-       wbinvd();
+       pmc_flush_cache();
 #endif
+       temp[2] = (temp[0] + 60) << 10;
+       temp[3] = (temp[2] + 60) << 10;
+       temp[4] = (temp[3] + 60) << 10;
+       temp[0] = (temp[4] + 60) << 10;
        //asm volatile ("");
+       barrier();
        __my_trace_mark(1, kernel_debug_test, NULL, "%d %d %ld %ld", 2, current->pid, arg, arg2);
+       barrier();
        //__my_trace_mark(0, kernel_debug_test, NULL, "%d %d %ld %ld", 2, current->pid, arg, arg2);
 }
 
@@ -69,11 +98,24 @@ static int my_open(struct inode *inode, struct file *file)
        unsigned long flags;
 
        local_irq_save(flags);
+#ifdef CACHEFLUSH
+       pmc_flush_cache();      /* initial write back, without cycle count */
+       msleep(20);     /* wait for L2 flush */
+#endif
        rdtsc_barrier();
        cycles1 = get_cycles();
        rdtsc_barrier();
        for(i=0; i<2000; i++) {
                test(i, i);
+               test(i, i);
+               test(i, i);
+               test(i, i);
+               test(i, i);
+               test(i, i);
+               test(i, i);
+               test(i, i);
+               test(i, i);
+               test(i, i);
        }
        rdtsc_barrier();
        cycles2 = get_cycles();